State-space model
Class of sequence models from classical control theory that compress a sequence into a latent state vector updated step by step — linear complexity in sequence length.
General form: h_t = A·h_{t-1} + B·x_t, y_t = C·h_t + D·x_t. The SSM revival in deep learning started with S4 (Gu et al. 2021), which showed that the right parameterisation of A matches transformer quality on Long Range Arena.
Mamba (2023) added input-dependent selectivity and became practical to scale. RWKV, Hyena and GLA are related families. Hybrids (Jamba, Zamba) mix SSM and attention blocks and seem to win on long context with good quality.